Exa gives AI applications web search and structured research

Exa supplies web search, page extraction and research infrastructure for applications and AI agents. A developer can request relevant pages, retrieve their content or delegate a longer research task that returns structured results. The practical distinction is control: Exa can supply evidence to your own application, or do more of the searching and assembling itself. Choosing between those routes affects response time, cost and how much of the evidence trail your team must maintain.

01 / ProductWhat Exa provides beyond a list of links

Exa’s Search API¹ accepts a query and supports constraints such as domains, publication dates and result categories. It can return page content alongside results. Those controls matter when the user needs a particular kind of evidence: an official product manual, for example, rather than a collection of pages that mention the same product. The API is an ingredient for a search-enabled application, not a guarantee that the application’s final answer is correct.

The Contents documentation² distinguishes page text, relevant passages and generated summaries. A passage preserves source wording; a summary adds another interpretation step. Our advice is to retain enough original context for the eventual reader to check an important claim. A compact response is useful for a model’s context window, but an extracted sentence can omit a qualification elsewhere on the page.

For broader tasks, Exa Agent³ accepts a research objective and an output structure, then runs asynchronously. Its documentation describes field-level grounding, run history and a cost breakdown. This shifts responsibility for orchestrating searches toward Exa. Your application still needs to decide whether a field is sufficiently supported and how to represent unknown or conflicting information.

02 / AudienceWho should use Exa, and who needs a different interface?

Exa is most directly relevant to developers building research assistants, company intelligence, documentation search and automated fact gathering. It is useful when a product needs repeatable retrieval with explicit inputs and outputs. A sales-research tool, for example, might need an official domain and supporting evidence for a company’s product category, rather than a persuasive paragraph about that company.

For a person who simply wants an interactive research answer, Perplexity’s user-facing workspace represents a different starting point. With an API, the developer owns the interface, storage and failure behaviour. With a finished research application, the reader works within that product’s modes and sharing model. Neither route is inherently superior; the question is whether you are buying an experience or building one.

Exa also warrants comparison with Tavily’s search and extraction approach. Compare them on your own question set, source coverage and integration needs. A generic search benchmark cannot tell you whether the provider finds an obscure documentation page your users depend on. Keep the final model and prompt fixed when comparing retrieval so that differences have a useful explanation.

03 / WorkflowA worked company-research workflow

Consider a proposed supplier directory that needs to establish whether a company offers a public API and where its pricing is documented. This is an illustrative architecture, not a benchmark we ran. Define the output before searching: company domain, API documentation URL, pricing URL, access date and a short evidence note. Include an explicit unknown state. Without it, a model may try to fill every field even when the public site does not answer the question.

Begin with the company’s known domain and search for the relevant documentation. Prefer a source controlled by that company for a claim about its own offering. Treat a similarly named business or an old reseller page as an entity-matching problem. The company name alone is a weak identifier; the domain and product context help keep results attached to the correct organisation.

Retrieve the promising pages and inspect the extraction result. Exa’s Contents guide² explains per-URL status reporting and configurable content views. A successful overall request does not mean every page yielded useful text. Preserve failed or empty extractions as such. For a current price, a page’s crawl date and the date on which the commercial terms apply answer different questions.

Turn the retrieved information into a structured record only after checking that it supports the field. A page headed Pricing may contain a sales form rather than a list price. Record that distinction directly. A useful final answer could say that enterprise pricing requires a quote and link to the page; it should not borrow a number from a cached blog post merely to make the table look complete.

When the task expands into discovering companies across several criteria, Agent³ is the documented route for multi-step assembly. Save the run identifier, wait for a terminal result and inspect the grounding supplied with each field. For a reader, show which details remain unverified. A returned JSON object is a convenient format; it is not independent verification of the business facts inside it.

04 / PricingExa pricing depends on the endpoint and result volume

The public pricing table⁴, checked on 15 September 2026, prices the following API operations in USD. These are usage prices, not monthly subscriptions. Search rates below cover requests with up to ten results; additional results and optional summaries can add charges. Enterprise terms and premium data providers need their own quote or rate check.

Illustrative arithmetic: at the listed Search rate, 10,000 requests within the base result allowance would cost $70 before extras, credits, taxes or contract adjustments. That is a calculation from the table, not a quote for a complete research service. If one user question triggers several searches, the number of user questions and the number of billable requests will differ. Model generation, storage and application hosting are additional parts of your own system.

The billing documentation⁵ says usage draws from team credits and distinguishes exhausted budgets from request-rate limits. Plan for both. A budget failure needs a spending decision or reduced work; a temporary rate limit needs controlled retry behaviour. Returning a friendly partial answer with sources already collected can be more useful than silently restarting the whole research task and spending again.

05 / DistinctionsSearch primitives and delegated research serve different jobs

The most useful design choice is how much of the retrieval loop to own. Direct calls give your application a clear place to apply source rules, stop conditions and caching. Delegated research is attractive when the task requires several searches and the desired output is already clear. In either case, define what evidence the interface exposes. A source link beside each important field makes the result easier to challenge and maintain than a bibliography attached to an undifferentiated answer.

Existing implementations should also notice the direction in Exa’s Websets guide⁶. It directs new list-building and enrichment workflows toward Agent while retaining instructions for existing Websets integrations, and says their credits are separate from Search API credits. A current evaluation should therefore identify the product actually being proposed, rather than assuming all Exa tutorials describe the same purchase and architecture.

06 / QuestionsQuestions to resolve before connecting production research

Can you trust the returned evidence?

Build an evaluation set with deliberately awkward cases: a renamed product, a discontinued service, a price requiring a billing selector and a company sharing its name with another organisation. Check whether the source genuinely supports the answer and whether the application declines to fill missing fields. Measure the work required to correct the output as well as the speed at which it arrives. We have not run this evaluation or independently measured Exa’s retrieval quality.

What leaves your application?

Search queries can reveal project names or customer interests even when the target material is public. Exa’s Enterprise pricing⁴ advertises zero data retention; confirm the contractual conditions and endpoint coverage before relying on that option. Separately minimise confidential context in search prompts. A public-web lookup for a supplier’s documentation rarely needs the private commercial reason you are evaluating that supplier.

Keep source content separate from application instructions. A retrieved page may contain irrelevant commands, sales language or malicious text. In the proposed directory, external content provides evidence about a company; it should not decide where the application sends data or which records it changes. Review the evidence-handling design before allowing a search agent to trigger actions in connected systems.

07 / DecisionChoose how much of the research loop to own